visual-slop-review: AI slop review for visual design

Design slop is the statistical average of the training corpus: safe, symmetric, purple, interchangeable. No single tell proves anything — several co-occurring tells mark a design as unintended. The knowledge base is self-contained in references/ (sources cited inline; measured research distinguished from designer folklore).

The cardinal rule (shared with ja-slop-review): breaking uniformity is a means, not a goal. Don't swap the AI template for an "anti-AI template" — every visual decision should trace to content, brand, or audience. And fixing 記号 (purple gradients, Inter) without fixing 構造 (meaningless hierarchy, decoration without purpose) leaves the design just as hollow.

When not to run this

Over-triggering costs more than under-triggering: 動いている UI に理由のない見た目の変更を入れるのは 退行で、「AI っぽい」判定を誤った文脈に当てると設計意図を壊す。Skip when:

• Functional-only work — バグ修正、ロジック変更、テスト、a11y 欠陥の修正。見た目を触らない変更に配色論を持ち込まない
• The user asked something else about the same screen — 挙動・パフォーマンス・文言・実装方針
• Throwaway artifacts — デバッグ用の図、自分用のスケッチ、社内の一時画面
• The brand guide or design system already decides it — 適合が要求なら、それに反する提案はしない (Process 1 と同旨)。トークンで決まっている色・書体を「AI っぽい」理由で変えない
• Charts as such — チャート・凡例・軸の設計は別スキル (dataviz) の領域。ここで見るのは周辺のレイアウトとタイポだけ
• The user already declined — 「今の見た目でいい」と言われた後の同一画面

Run at most once per artifact per conversation. 一度通した画面に後続ターンで tell 探しを重ねない — 尽きた後の追加指摘は、根拠のない差分に落ちる。

A caveat is the tell — 「これは LP 向けの基準だが、業務 UI にも役立つかも」と前置きしたくなった 時点で、その指摘は当たっていない (references/ の抽出元は LP / マーケティングサイト中心)。

Process (run in order)

1. Identify the artifact type and its budget — web UI / slide deck / diagram / image, plus the governing brand or design system. If a brand guide or design system exists, conformance to it outranks every technique below.
2. Scan the tell catalogue for the type:
   • Web UI: references/web-ui.md (color/gradient, typography, layout, motion, UI copy)
   • Slides & diagrams: references/slides-diagrams.md (uniform bullets, rigid layout, 図の文法エラー)
   • Images & illustration: references/images.md (durable vs dated tells — anatomy tells are decaying, physics tells persist)
3. Separate 記号 from 構造 — list symbol-level fixes (palette, fonts, icons) and structure-level fixes (hierarchy matching importance, relationships made explicit, content-driven layout) independently. Both are required.
4. Apply references/prescriptions.md — constraints before generation (tokens, negative constraints, WCAG), content-connected decisions during design, separated evaluation after (5-second / first-click / one-variable comparison).
5. Hand text to the sibling skill — UI copy, slide prose, and Japanese body text go through ja-slop-review; the two skills catch the same slop from the visual and verbal sides.

Quick checklist

• Palette/fonts chosen for a reason traceable to brand or content — not indigo/purple-blue gradient, cream-beige + terracotta (#D97757), near-black + a single acid-green / vermilion accent, tinted near-black (#0B0B0B / #111), or Inter-by-default. Text contrast ≥ 4.5:1 (3:1 for large text)
• Real typographic hierarchy (role-based pairing, size + weight + spacing), not one family with weight changes; no single accented word in the headline (serif italic, bold, or another color), no monospace or tiny-uppercase eyebrow labels by default; body line length under ~80 characters
• No identical icon-tile feature-card grids, badge-pill or pill-button reflexes, numbered "01/02/03" section labels that encode no real order, cardocalypse, broadsheet look (hairline rules, zero radius, dense columns) by default, template chrome (A · B · C meta strings, WORD — fragment labels, a trailing → on links and buttons), or uniform radius/shadow everywhere; layout derived from actual content, visual weight matches importance
• Icons: one family, adopted for a reason; no emoji-as-icons; ✨ never means "AI" without a label
• Motion explains state, continuity, or attention — otherwise deleted
• Slides: type/color/icons unified across the deck, layout varies only at meaningful transitions; bullet rhythm broken on purpose
• Diagrams: one relationship type chosen (parallel / causal / sequence / comparison); every arrow and label verified; built as editable vector, not a single raster
• Images: default AI look actively steered away (art-direction prompt + references); no Japanese text baked in; light/shadow/reflection consistent
• Symbol fixes accompanied by structure fixes — if only the colors changed, it is still slop

鮮度

この知識ベースは出所リポジトリの clone を持たない自作の調査資料なので、SHA ではなく日付で棚卸しする。各 references/*.md の frontmatter が tracking: review + reviewed_at + review_interval_days を持ち (images.md / web-ui.md は 90 日 — 生成モデルの世代と配色トレンドに依存するため、他は 180 日)、/yokai-refs-update が期限切れを REVIEW として報告する。とくに images.md の「陳腐化」ラベル付き tell は世代交代で無効化されるので、棚卸しのたびに持続性ラベルを見直す。